Moonshot AI this week unveiled its latest large language model, Kimi K3, claiming it packs 2.8 trillion parameters and matches the benchmark performance of models from OpenAI and Anthropic. The announcement positions the Beijing-based startup as a serious contender in the global AI race, though independent verification of the claims is still pending.
A 2.8-trillion-parameter model
Kimi K3’s parameter count puts it among the largest models ever disclosed. For context, OpenAI’s GPT-4 is widely believed to have around 1.7 trillion parameters, though neither company has officially confirmed the figure. Anthropic’s Claude models are also estimated to be in the hundreds of billions to low trillions. Moonshot AI didn’t release a technical paper or third-party benchmarks alongside the announcement, so the numbers are purely the company’s own.
Benchmark performance claims
The company said Kimi K3 performs on par with OpenAI and Anthropic models across a range of standard AI benchmarks. That’s a strong claim, given that those two firms have dominated the space for years. Moonshot AI has not specified which benchmarks were used or provided comparative scores. Outside researchers will likely need to run their own tests before the claim gains widespread acceptance.
What’s next
Moonshot AI said it plans to release an API for Kimi K3 in the coming months, but hasn’t shared a specific date. The company also hasn’t disclosed pricing or availability for developers outside China. Until then, the AI community will be watching for independent evaluations — and for how OpenAI and Anthropic respond to a new challenger with a very large model.
